sumcheck protocol support mixed num_vars monomial form

[hero78119]

Goal

To make sumcheck protocol support different num_vars, aiming for

• minimal change to sumcheck protocol, make verifier remain the same. no extra meta data passed from prover, and what prover have just mle and it's eval size

Besides this PR also remove some parallism in verifier since it's unnecessary for relative low cost.

design rationale

(Also comments in codebase for reference)
To deal with different num_vars, we exploit a fact that for each product which num_vars < max_num_vars,
for it evaluation value we need to times 2^(max_num_vars - num_vars)
E.g. Giving multivariate poly $f(X) = f_1(X1) + f_2(X), X1 \in {F}^{n'}, X \in {F}^{n}, |X1| := n', |X| = n, n' &lt;= n$
For i round univariate poly, $f^i(x)$
$f^i[0] = \sum_b f(r, 0, b), b \in [0, 1]^{n-i-1}, r \in {F}^{n-i-1}$ chanllenge get from prev rounds
= $\sum_b f_1(r, 0, b1) + f_2(r, 0, b), |b| &gt;= |b1|, |b| - |b1| = n - n'$
= $2^{(|b| - |b1|)} * \sum_{b1} f_1(r, 0, b1) + \sum_b f_2(r, 0, b)$
same applied on f^i[1]
It imply that, for every evals in f_1, to compute univariate poly, we just need to times a factor 2^(|b| - |b1|) for it evaluation value

benchmark

benchmark with ceno_zkvm riscv_add, and gkr keccak both remain the same and no impact.
You might see some redundancy coding style, but this is for retain the best performance. I tried other variants and it impact benchmark results

scope

Related to #109 #210 ....
To address #126 #127
This enhance protocol features potiential can be used for range table-circuit, init/final-memory, cpu-init/cpu-final halt to make selector sumcheck support batching different num_instance witin.

[hero78119]

@kunxian-xia I plan to study what the minimal change to batching multiple table (or similar) circuit in single create_table_proof after this PR, since each table circuit only contribute one wit-in. The previous blocker for batching them is on selector sumcheck which require on same num_instance (whereas tower protocol already support)
So after this features it's possible for us to batch those naive constrains system into the same sumcheck proof

[hero78119]

The idea behind this is for skip f.len() == 1 assuring we dont need to add extra boundary check in line 455/456 so no performance impact

[dreamATD]

LGTM!